Wang Qifeng frequently visits district and county-level hospitals to provide clinical guidance.
As a chief physician in the radiotherapy department at Sichuan Cancer Hospital, he has encountered numerous regrettable cases at the grassroots level. Patients often arrived unable to eat, having already missed the window for curative treatment. When he pulled up their hospital records, he would find that the same person had been hospitalized for pneumonia two years earlier and had undergone a chest CT scan.
Comparing that scan from two years ago with the current one, the esophageal wall had already shown thickening at that time.
No one noticed it back then. The patient had not mentioned any difficulty eating, and the radiologist had only examined the lung condition.
On September 24, Alibaba's DAMO Academy, in collaboration with Sichuan Cancer Hospital, Sun Yat-sen University Cancer Center, and other institutions, released an AI model for esophageal cancer screening called DAMO EAGLE. The related paper was published in Nature Medicine. During the initial screening phase, the model directly reads plain CT scans without requiring intubation or contrast agent injection. After the AI flags a positive result, patients must still undergo endoscopy and pathological examination for a definitive diagnosis.
The DAMO EAGLE paper was published in Nature Medicine.
What it aims to do is help doctors examine the esophagus that gets captured incidentally during lung scans.
This is the fourth cancer screening model released by DAMO Academy, following those for pancreatic cancer, gastric cancer, and colorectal cancer. Last year's gastric cancer model, GRAPE, was designed for use before gastroscopy. EAGLE, by contrast, starts from chest CT scans that hospitals have already taken to search for esophageal cancer risk.
Gold standard exists, but too few receive endoscopy
China's esophageal cancer burden stands out globally. Data from the National Cancer Center shows that in 2024, China recorded approximately 200,000 new cases of esophageal cancer and about 172,000 deaths, accounting for nearly half of the global total.
China's major cancer deaths in 2024 (in ten thousands), with esophageal cancer at 172,100.
High-incidence areas are concentrated in the Taihang Mountain region spanning Henan, Hebei, and Shanxi provinces, as well as the Huai River basin, Sichuan, Guangdong, and Fujian. Rural areas have significantly higher rates than urban areas. Wang Qifeng noted that this is related to local dietary habits of consuming hot food and hot tea.
About 70% of patients are already at an advanced stage when diagnosed.
The prognosis difference between early and late stages is enormous. When the lesion is still confined to the mucosal layer, a single minimally invasive endoscopic procedure can achieve a cure, with a five-year survival rate exceeding 95%. Once metastasis occurs, Wang Qifeng cited figures showing that regardless of the treatment method used, the five-year survival rate is only about 10% to 15%.
From the initial sensation of a foreign body when swallowing to being able to consume only liquids, the progression typically takes just about six months.
Wang Qifeng has a personal memory of this process. When he was a second-year medical student, his grandmother was diagnosed with esophageal cancer in Changzhi, Shanxi, which is also a high-incidence area. At diagnosis, the lesion was already about 5 centimeters. After the first phase of treatment, his grandmother regained the ability to eat, but eventually developed an esophageal fistula. Only eight months passed from diagnosis to her death.
Methods for early detection do exist. Upper gastrointestinal endoscopy is the recognized gold standard. The problem lies in the fact that there simply are not enough resources to perform them all.
According to official estimates, China has over 100 million people at high risk for esophageal cancer who need screening, yet only about 30 million upper gastrointestinal endoscopies are actually performed each year.
The shortfall stems from several factors. Training an endoscopist takes three to five years, and a single examination takes 20 to 30 minutes. A standard endoscopy costs two to three hundred yuan, while the painless version that most people prefer costs around 1,000 yuan. Then there is fear. A tube about one centimeter in diameter goes from the mouth down to the stomach. Wang Qifeng said that even his own parents and relatives are afraid of it.
High-incidence areas also tend to be places with the fewest endoscopists.
The hardest organ, difficult because even doctors cannot annotate it
Plain CT is one of the most commonly used imaging examinations in China, performed in outpatient clinics, emergency rooms, health checkups, and inpatient settings, generating hundreds of millions of images each year. The scanning range of a chest CT inherently covers the esophagus.
But the industry has generally believed that finding early-stage esophageal cancer with plain CT is nearly impossible.
The esophagus is a muscular tube about 25 centimeters long. It is normally closed and constantly moves when speaking or swallowing saliva. It sits right next to the heart and major blood vessels, and breathing and heartbeat both alter its shape. Early lesions are confined to the epithelial or mucosal layer, producing very minimal changes in thickness and density on CT.
Guo Guangyu, an algorithm expert at DAMO Academy, offered an analogy: it is like finding a small fuzzball on a wrinkled pant leg.
Wang Qifeng compared it with the pancreas. The pancreas is a solid organ, so even a small abnormality may produce a visible signal on imaging. The esophagus leaves only a fleeting impression on CT, and changes must exceed 5 millimeters before they are judged as abnormal.
The very first step of model training, annotation, proved extremely difficult. Doctors cannot clearly see early lesions on plain CT, so there is no way to begin annotating training data.
The team's solution was to use other examinations for localization. They first used endoscopic surgery and pathology reports to determine the centimeter-level location of the lesion. Then, specialized esophageal radiologists annotated it precisely on contrast-enhanced CT. Through image registration, these annotations were "transferred" onto the same patient's plain CT. Each case was annotated independently by two doctors, and when their opinions differed, a doctor with eight years of experience reviewed the case.
The second problem involved low-dose CT. Health checkups for lung nodules mostly use low-dose CT, but low-dose CT scans from esophageal cancer patients are very hard to accumulate. Guo Guangyu explained that once a patient discovers an abnormality, the CT they receive at the hospital is a standard-dose scan.
The team therefore built a "virtual CT machine," feeding standard-dose CT scans into it to reduce the dose and simulate low-dose images for training the model.
The model works in two steps. First, it segments the esophageal region from the entire CT scan to eliminate interference from surrounding structures. Then it performs a detailed analysis of the esophagus, simultaneously providing a patient-level benign or malignant judgment, lesion contours, and a heatmap. Doctors can see where the AI is looking.
DAMO EAGLE identifies a T1-stage esophageal cancer lesion on plain CT. From left to right: plain CT, physician annotation, model segmentation result, and heatmap.
In early 2024, the model underwent its first large-scale real-world retrospective study, achieving a sensitivity of about 85% and specificity of approximately 98%.
The current version has been validated across 3 countries, 12 centers, and more than 80,000 patients. According to the paper, in opportunistic screening scenarios, the model achieved a sensitivity of nearly 90% for esophageal cancer and 52.5% for precancerous lesions, which produce even weaker signals. After real-world iteration, the specificity in clinical settings reached 99.2%.
The low-dose version achieved a sensitivity of 88.4%, essentially comparable to the standard-dose version. Among 10,959 asymptomatic health checkup participants, the specificity was 99.94%.
Regarding the specificity figure, Guo Guangyu offered a direct conversion: "One false positive means a patient may have to spend over 1,000 yuan more on an additional examination."
A tool integrated into the health checkup workflow faces mostly healthy individuals, and the false positive rate determines whether it can truly be put into practical use.
Not enough endoscopes, so let CT pick the people first
Wang Qifeng was very clear about EAGLE's positioning: "AI is not meant to replace endoscopy, but to screen out high-risk individuals who then undergo endoscopy. This makes endoscopy more precise and greatly improves the detection rate."
The research team validated EAGLE separately in high-risk population screening and routine hospital CT scans.
First, high-risk populations. Participants undergo a plain CT scan, and EAGLE runs in high-sensitivity mode. Those flagged as positive are prioritized for endoscopy.
In July 2023, the team conducted a simulated "CT initial screening plus endoscopic confirmation" study in Suining, Sichuan. Results showed that the detection rate of malignant lesions by endoscopy increased from 1.7% to 5.2%, and the number of endoscopies needed to detect one malignant lesion dropped from 59 to 19.
Different uses of EAGLE in routine hospital CT and gastroscopy screening for high-risk populations. Redrawn based on data from the Nature Medicine paper.
These are not complete prospective clinical results. The 3 esophageal cancer cases in the mixed cohort came from retrospective case resampling. How much endoscopy can be reduced still needs to be verified in real screening programs.
Another validation group used chest CT scans that hospitals had already taken. Lung cancer screening already uses low-dose chest CT, and hospitals also have large numbers of patients who undergo chest CT for pneumonia, cough, and inpatient examinations. EAGLE looks for esophageal abnormalities on these images without adding an extra scan. After the AI flags a positive result, doctors then determine whether gastroscopy is needed.
The pneumonia CT scan from two years ago that Wang Qifeng encountered falls into this category.
High-risk screening and routine hospital CT place different demands on the model. The former prioritizes minimizing missed diagnoses, while the latter faces mostly people without esophageal cancer, where fewer false positives mean less burden on doctors for review and on patients for follow-up. Wang Qifeng said the model's threshold can be adjusted by scenario.
For high-incidence areas, CT scanners already widely available at the grassroots level can take on initial screening, without needing to build a separate imaging infrastructure.
The plain CT path, now at its fourth cancer type
DAMO Academy began researching "plain CT plus AI" multi-cancer screening in 2021, successively releasing models for pancreatic cancer, gastric cancer, colorectal cancer, and esophageal cancer, as well as covering major diseases such as aortic dissection. According to DAMO Academy's disclosures, related research has been published in Nature Medicine five times cumulatively. Among these, the pancreatic cancer model PANDA received FDA "Breakthrough Device" designation in April 2025, and this technical approach has also been adopted by the national major science and technology project for the "Four Major Chronic Diseases."
Overview of DAMO Academy's medical AI models (Source: DAMO Academy).
The EAGLE project was initiated in 2022. The opportunity that made Wang Qifeng willing to participate was PANDA. His thinking at the time was simple: every day he treats advanced-stage patients, the treatment process is painful, and survival is not long. If a model could be trained from the CT scans taken every day to detect early-stage lesions, it could save more lives.
Multi-cancer early screening is also exploring liquid biopsy. Take Grail's Galleri as an example: participants must have blood drawn separately, and cancer clues are sought from cfDNA methylation signals. EAGLE analyzes CT scans that hospitals have already produced, eliminating the need for an additional initial screening test. Confirmation still requires subsequent specialty procedures.
Chest CT and abdominal CT cover different organs. After the AI flags a positive result, patients must still undergo endoscopy, pathology, or other specialty examinations.
Hospitals can reuse existing images but must also take on more review, referral, and follow-up work.
According to DAMO Academy's vision, existing chest or abdominal plain CT scans at hospitals could in the future be extended to cover lung cancer, liver cancer, colorectal cancer, gastric cancer, esophageal cancer, pancreatic cancer, and breast cancer. Because scanning ranges differ, a single CT cannot simultaneously cover all seven cancer types, but the same imaging infrastructure can be reused.
EAGLE still has some way to go before large-scale use. The model's sensitivity for high-grade intraepithelial neoplasia is 52.5%, meaning nearly half of these cases are not identified.
In prospective hospital validation, 42 of 90 AI-positive individuals did not enter the existing diagnostic and treatment workflow. A multidisciplinary team recommended that 16 of them undergo gastroscopy, but ultimately only 3 completed the examination, and 1 case of esophageal cancer was found. After the AI reports a positive result, hospitals still need to arrange imaging review, specialty referral, gastroscopy, and ongoing follow-up.
Gastroscopy completion rates, the number of newly detected early-stage cases, the cost of false-positive follow-up examinations, and whether esophageal resection, chemoradiotherapy, and late-stage treatment can be reduced all still need to be answered in larger-scale prospective studies.
EAGLE has already validated a feasible direction: using the plain CT scans generated by routine hospital examinations to screen for high-risk individuals with solid tumors, going beyond the original diagnostic purpose of those scans.
Comments